• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/scalar-rr2-p5-div.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 #include <math.h>
12 
13 #include <xnnpack/common.h>
14 #include <xnnpack/vunary.h>
15 
16 #include <fp16/bitcasts.h>
17 
18 
xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x4(
20     size_t n,
21     const float* x,
22     float* y,
23     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(n % sizeof(float) == 0);
26 
27   const float vmagic_bias = params->scalar_rr2_p5.magic_bias;
28   const float vminus_log2e = params->scalar_rr2_p5.minus_log2e;
29   const float vln2_hi = params->scalar_rr2_p5.ln2_hi;
30   const float vln2_lo = params->scalar_rr2_p5.ln2_lo;
31   const float vc5 = params->scalar_rr2_p5.c5;
32   const float vc4 = params->scalar_rr2_p5.c4;
33   const float vc3 = params->scalar_rr2_p5.c3;
34   const float vc2 = params->scalar_rr2_p5.c2;
35   const float vc1 = params->scalar_rr2_p5.c1;
36   const float vone = params->scalar_rr2_p5.one;
37   const float vdenorm_cutoff = params->scalar_rr2_p5.denorm_cutoff;
38 
39   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
40     const float vx0 = x[0];
41     const float vx1 = x[1];
42     const float vx2 = x[2];
43     const float vx3 = x[3];
44     x += 4;
45 
46     const float vz0 = fabsf(vx0);
47     const float vz1 = fabsf(vx1);
48     const float vz2 = fabsf(vx2);
49     const float vz3 = fabsf(vx3);
50 
51     float vn0 = vz0 * vminus_log2e + vmagic_bias;
52     float vn1 = vz1 * vminus_log2e + vmagic_bias;
53     float vn2 = vz2 * vminus_log2e + vmagic_bias;
54     float vn3 = vz3 * vminus_log2e + vmagic_bias;
55 
56     const float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23);
57     const float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23);
58     const float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23);
59     const float vs3 = fp32_from_bits(fp32_to_bits(vn3) << 23);
60 
61     vn0 -= vmagic_bias;
62     vn1 -= vmagic_bias;
63     vn2 -= vmagic_bias;
64     vn3 -= vmagic_bias;
65 
66     float vt0 = vn0 * vln2_hi + vz0;
67     float vt1 = vn1 * vln2_hi + vz1;
68     float vt2 = vn2 * vln2_hi + vz2;
69     float vt3 = vn3 * vln2_hi + vz3;
70 
71     vt0 = vn0 * vln2_lo + vt0;
72     vt1 = vn1 * vln2_lo + vt1;
73     vt2 = vn2 * vln2_lo + vt2;
74     vt3 = vn3 * vln2_lo + vt3;
75 
76     float vp0 = vt0 * vc5 + vc4;
77     float vp1 = vt1 * vc5 + vc4;
78     float vp2 = vt2 * vc5 + vc4;
79     float vp3 = vt3 * vc5 + vc4;
80 
81     vp0 = vt0 * vp0 + vc3;
82     vp1 = vt1 * vp1 + vc3;
83     vp2 = vt2 * vp2 + vc3;
84     vp3 = vt3 * vp3 + vc3;
85 
86     vp0 = vt0 * vp0 + vc2;
87     vp1 = vt1 * vp1 + vc2;
88     vp2 = vt2 * vp2 + vc2;
89     vp3 = vt3 * vp3 + vc2;
90 
91     vp0 = vt0 * vp0 + vc1;
92     vp1 = vt1 * vp1 + vc1;
93     vp2 = vt2 * vp2 + vc1;
94     vp3 = vt3 * vp3 + vc1;
95 
96     vt0 *= vs0;
97     vt1 *= vs1;
98     vt2 *= vs2;
99     vt3 *= vs3;
100 
101     const float ve0 = vt0 * vp0 + vs0;
102     const float ve1 = vt1 * vp1 + vs1;
103     const float ve2 = vt2 * vp2 + vs2;
104     const float ve3 = vt3 * vp3 + vs3;
105 
106     const float vd0 = ve0 + vone;
107     const float vd1 = ve1 + vone;
108     const float vd2 = ve2 + vone;
109     const float vd3 = ve3 + vone;
110 
111     float vf0 = ve0 / vd0;
112     float vf1 = ve1 / vd1;
113     float vf2 = ve2 / vd2;
114     float vf3 = ve3 / vd3;
115 
116     if XNN_UNPREDICTABLE(vz0 > vdenorm_cutoff) {
117       vf0 = 0.0f;
118     }
119     if XNN_UNPREDICTABLE(vz1 > vdenorm_cutoff) {
120       vf1 = 0.0f;
121     }
122     if XNN_UNPREDICTABLE(vz2 > vdenorm_cutoff) {
123       vf2 = 0.0f;
124     }
125     if XNN_UNPREDICTABLE(vz3 > vdenorm_cutoff) {
126       vf3 = 0.0f;
127     }
128 
129     if XNN_UNPREDICTABLE(vx0 > 0.0f) {
130       vf0 = vone - vf0;
131     }
132     if XNN_UNPREDICTABLE(vx1 > 0.0f) {
133       vf1 = vone - vf1;
134     }
135     if XNN_UNPREDICTABLE(vx2 > 0.0f) {
136       vf2 = vone - vf2;
137     }
138     if XNN_UNPREDICTABLE(vx3 > 0.0f) {
139       vf3 = vone - vf3;
140     }
141 
142     y[0] = vf0;
143     y[1] = vf1;
144     y[2] = vf2;
145     y[3] = vf3;
146     y += 4;
147   }
148   if XNN_UNLIKELY(n != 0) {
149     do {
150       const float vx = *x++;
151 
152       const float vz = fabsf(vx);
153 
154       float vn = vz * vminus_log2e + vmagic_bias;
155       const float vs = fp32_from_bits(fp32_to_bits(vn) << 23);
156       vn -= vmagic_bias;
157 
158       float vt = vn * vln2_hi + vz;
159       vt = vn * vln2_lo + vt;
160 
161       float vp = vt * vc5 + vc4;
162       vp = vt * vp + vc3;
163       vp = vt * vp + vc2;
164       vp = vt * vp + vc1;
165 
166       vt *= vs;
167       const float ve = vt * vp + vs;
168       const float vd = ve + vone;
169 
170       float vf = ve / vd;
171       if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
172         vf = 0.0f;
173       }
174       if XNN_UNPREDICTABLE(vx > 0.0f) {
175         vf = vone - vf;
176       }
177 
178       *y++ = vf;
179 
180       n -= sizeof(float);
181     } while (n != 0);
182   }
183 }
184